Email Calvin || Email Bickerton || Book's Table of Contents || Calvin Home Page || September 1999


William H. Calvin and Derek Bickerton, Lingua ex Machina: Reconciling Darwin and Chomsky with the human brain (MIT Press, 2000), chapter 8.  See also

copyright ©2000 by William H. Calvin and Derek Bickerton

The nonvirtual book is available from or direct from MIT Press.

Webbed Reprint Collection
This 'tree' is really a pyramidal neuron of cerebral cortex.  The axon exiting at bottom goes long distances, eventually splitting up into 10,000 small branchlets to make synapses with other brain cells.
William H. Calvin

University of Washington
Seattle WA 98195-1800 USA


A Common Code:
The Brain's "Esperanto" Problem

Many of our concepts have multiple sensory modalities associated with them. The flowers in the formal gardens that cascade below the Villa Serbelloni have, in my head, both a visual category and an associated smell (and often an associated insect!), from all my trips up the path.

If the concept is a word, it also has some associated movements, needed to pronounce or write the word. Neocortex is where the sight of a comb, say, is matched up to the feel of a comb in your hand. While the spatiotemporal firing patterns for the comb=s sight and feel are likely different, they become to be associated in the cortex B along with those for hearing the sound /km/ or hearing the characteristic sounds that the teeth of a comb make when they=re plucked. Any of these inputs could enable you to say, AThat=s a comb.@ On the production side, you not only have linked spatiotemporal patterns for pronouncing /km/ but also ones for manipulating a comb through the hair on your head, or for writing down the word on your list of left-behind items that need to be replaced on a walk down into the streets of Bellagio. There are likely a dozen different cortical codes associated with combs, all located (as Sigmund Freud suspected a century ago) in different places in the cortex. How can we link them? Say, link the concepts of the temporal lobe with the verbs of the frontal lobe?

The largest bundle of nerve fibers in the brain is the corpus callosum B which, as every intro psych student knows, connects the right brain with the left brain. But the second largest is the arcuate fasciculus, connecting the temporal lobe with the frontal lobe on the same side. The left arcuate fasciculus has got to be heavily involved in communicating temporal lobe concepts to any sentence-planning machinery in the frontal lobe part of the language system. There are subcortical paths involving chains of neurons in the thalamus or basal ganglia that also connect the two lobes, but the arcuate fasciculus is largely made up of the direct corticocortical paths; they=re branches of the same sideways axons whose express train patterning created all the interesting possibilities for cloning spatiotemporal patterns.

The arcuate fasciculus is a mere pipeline, analogous to a fiber optic bundle of thousands of thin light pipes. It is not like the better kinds of fiber optic bundles, the coherent type that they use for endoscopes. When neighboring fibers at one end aren=t still neighbors by the time they reach the far end, having gotten jumbled somewhere along the way, such bundles are called Aincoherent.@ Used for viewing some internal organ, displacements would occur, rather like the way Picasso put eyes in the middle of foreheads. Even when nothing as dramatic as jumble happens in our cerebral pathways, things are still blurred because each fiber fans out at the far end, spanning a millimeter or more.

What all this incoherence means is that the firing pattern within a neural committee (what Hebb called a cell-assembly, and what I claim can be contained in a half-millimeter hexagonal patch of cortex) cannot be communicated undistorted to another region of the brain B not the way it could if the arcuate fasciculus were as coherent as the fiber optics in an endoscope. I=m not going to propose that the overall size of the arcuate fasciculus has changed disproportionately, only that the Achoirs@ that drive it on occasion finally become large enough to achieve coherent transmission, perhaps because the Darwinian competition in the sending cortex grew a particularly extensive hexagonal mosaic.

Fortunately, the jumble and blur aren=t really a problem most of the time. Though incoherent projections are probably the natural state of affairs in all animals, and choirs are likely lacking, there is a simpler way of handling everyday messages. The codes are spatiotemporal firing patterns, as abstract as a bar code, and the distorted version, f(A), reaching the far end will serve equally as well as A (the code space is so enormous that there is little chance of the distorted code hitting upon a code already in use B though, when it happens, you might confuse the sight of an apple with the smell of an orange B so-called synesthesia).

Incoherence is a problem, however, when the receiving cortex wants to send the distorted code f(A) back to the originating cortical area (always via a separate
Aroad@ within the arcuate fasciculus, as any given axon is a one-way street). Six out of seven cortical areas have such reciprocal connections. So the apple code arrives back at its origin, doubly distorted into g[f(A)] B not the same thing at all. But no matter, you might reasonably claim: the originating cortex is smart enough so it can eventually learn, with enough repetitions, that the spatiotemporal pattern g[f(A)] means the same thing as A. If A was the code for an apple, f(A) and g[f(A)] are alternative codes B it=s rather like apple, apfel, and mela (saved by my Italian phrase book) all mean the same thing but in different places. With time spent learning German, apple and apfel bring up the same association B and so it should be with distorted spatiotemporal firing patterns, provided there is time to do all the learning, and provided that the originating cortex can perform this trick with each of the dozens of cortical areas with which it maintains a correspondence (and that means something like N-factorial identities have to be gradually established).

There are some real advantages of a common language, a universal code for apple that would work in all places B a sort of Esperanto apple in the brain=s various places concerned with apple taste, shape, smell, pronunciation, and so forth. Such a common code would allow novel associations to be formed on the fly, rather than through a laborious pairing procedure. Language is full of never-seen-before associations, such as Aa square blue tomato,@ that you have to work with, passing it around for awhile until some area tentatively responds to it.

DB: This is great! I can see now why protolanguage is so different from true language. When words go to the motor organs of speech one at a time, as they do in protolanguage, lack of coherence isn=t a problem B you don=t have to join them with other words before you utter them, and you don=t have to pass them back and forth. But if words have to be tentatively assembled before uttering them, an incoherent system would change them during the assembly process so that what came out would be just word salad. A common code is an all or nothing thing B you either have one or you don=t.

But a common code for apple in even two cortical areas requires coherence in those long corticocortical bundles of axons B which I just said was unlikely on the basis of the typical jumble and fanout in the observed neuroanatomy.

Ah, but anatomy is not destiny B there is a physiological way to recreate the original undistorted pattern, A, in the receiving area. And that physiological Good Trick provides a second candidate for the Great Leap Forward, the evolutionary Good Trick allowing syntax to emerge from protolanguage (or, if social-calculus-based argument structure got there first, allowing a great expansion of the workspace that would support many clauses and phrases). Creating order out of incoherence has a critical mass involving a redundancy, not unlike the error-correcting codes used on computer disks.

Given that each spatiotemporal firing pattern is like a little tune, what my good trick requires is simply a plainchant chorus, all singing the same song. The internal synchronization is provided by the aforementioned axon of the superficial pyramidal neuron, the one whose terminal fanout occurs in a nested set of concentric bands, rather as an express train skips intermediate stops. When the chorus is large enough, and they=re all sending axons to the target cortex, you can actually recreate several adjacent instances of code A despite the jumble and blur B provided (and this is a theoretical assumption at this time, not data; see chapter 8 of Code) that the axon=s fanout in the receiving cortex is also a nested set of concentric bands, much like the other axon branches are known to produce in neighboring regions of the sending cortex.

You don=t (in my theory) recreate the whole redundant pattern from the sending cortex at first, the way it originally progressed. All it takes is an adjacent pair of hexagons, each containing the unit pattern, somewhere in the jumbled and blurred projection. That pair then seeds the proper melody for what might become a sizeable chorus of A in the distant receiving cortical area, a hexagonal mosaic that spreads sideways and recruits new territory. If there is a good resonance for that spatiotemporal pattern in the receiving cortex, the resulting chorus might become even larger than the one that started it from the sending cortex (much like metastatic tumors seed themselves in distant organs).

For any given degree of jumble and blur, there is a critical mass for the sending choir, a number of simultaneous singers (hexagons in the mosaic) below which the receiving cortex cannot reproduce the exact spatiotemporal pattern, not even in a minimal pair of hexagons. And so the distorted spatiotemporal firing pattern, code f(A), has to be used instead, with all its attendant disadvantages for on the fly associations.

Resonances to A could be in the sending cortex alone, with the receiving cortex merely repeating it. Or the sending cortex might be a mere sensory buffer without long-term memory at all; it would send out A to various distant areas, asking their resonances, in effect, AAnyone recognize this?@ I did mention, didn=t I, that the visual attributes of an apple are likely to reside near visual cortex, that its pronunciation resonance is likely to be near auditory cortex, and so forth?

The cortical area that has a resonance will promptly form up a large chorus singing the song of A, activating similar links to other areas. A successful resonance in one area could trigger the request of the whole cortex=s distributed data base; those resonances might contribute the smell of A or the typical pronunciation of A. If all of the relevant regions form up local choruses of A, you get a particularly evocative apple and the word is on the tip of your tongue.

So the concept of an apple is not stored in some particular location; it=s best thought of as its universal cerebral code, the spatiotemporal firing pattern A. There are likely local codes as well, Av, Ap, and so forth, each different because of the contingent history of how we came to learn the sight of an apple, how we learned to pronounce it, and so forth. Each will require an identity with our Esperanto A, the common code for long-distance communication within cortex. At least we no longer need N! identity associations as with incoherent corticocorticals, each slowly learned before becoming useful. The common code A is, of course, equally idiosyncratic B my apple code is surely different than yours B and, I suppose, it=s probably closely related to one of the local codes, Av, Ap, and so forth, perhaps the little tune that first achieved widespread use in other cortical areas.

The nice bonus from this common code is not only efficiency but provisional associations: we can now form novel associations on the fly, as when we first encounter an Italian version of apple pie, proceed to store a resonance for the combination of apple code and pie code, and next trigger it when we again encounter the dish, complete with the link to Villa Serbelloni, identifying where we first tasted it. We can even imagine such a dish, and store the imagined associations.

This is, obviously, not the most elementary way of doing one-trial associative memories; I=m sure that snails learning food avoidance manage without hexagonal mosaics of Hebbian cell-assemblies. But behavior has a time constraint: there are fleeting windows of opportunity. You have to survey the possibilities in a limited period of time and, as brains get bigger and contain more information, that access time may lengthen. Having a common code means you can take alternative paths through the brain to wherever the resonances actually reside in the cortex; there=s no path dependence from a series of learned translations of codes, those pesky identities. Just think of the tortuous path that the ancient Greek classics took: Greek to Arabic to Latin to vernacular, with dropouts all along the way because no one was sufficiently interested in preserving or translating a work. That=s likely what happens in the neocortex, without a common code from which one-step translations can be done into each modality=s working code. When my books are translated, the Hungarian version isn=t done from the German, which was done from the Dutch, which was the one done from the original English. Rather than seriatim, they=re all done in parallel from a common American English source.

In the case of cerebral sensory and movement modalities, a common code means that a number of routes can be taken from visual to motor cortex, not just one slow route of translation after translation after translation. Each area can translate from the original common code into its local scheme, each can send on the message in the common code B and that=s what makes things so flexible and so much better able to handle the novel links needed for the juxtapositions that language tasks utilize.

The common code is often a superposition of codes, with the local area occasionally tacking on another attribute generated from its local resonances. That, Derek, is what I take to be your many-faceted word. A multivoiced symphonic version of the hexagonal code can be a sentence, but let me save that for later, after you discuss the big step up.



DB: Sounds to me like messages in the brain that are not fully coherent are not partially coherent, they are completely incoherent, and with anything less than a coherent message, the brain could not produce a sentence. Is that where you=re going?

WHC: Not quite. I=m only saying that the usual incoherent corticocorticals aren=t as useful for doing novel associations on the fly. Lacking the speed from corticocortical coherence, you=d make your multiregional connections far too late to be useful for behavioral windows of opportunity.

Recall that I said that a sentence is not just a heap of words, not any more than a house is a mere heap of construction materials? Well, the sentence is really a house of cards, and the breeze that can topple it is called incoherence.

Perhaps I should explain the many uses of the coherence concept, as the term comes with a lot of baggage. (E.O. Wilson also despaired of this and resurrected an old synonym, consilience, for his grand book on how science fits together.) Coherence, outside of the fiber-optics technical context I=ve been using so far, just means the logical, consistent, orderly relation of various parts. When you use the term more broadly, you=re implying that various aspects fit together well, that it Aall hangs together@ nicely. I might speak of an incoherent memory if I=d confused two people, remembering the face of one with the foreign accent of another.

Incoherence often happens during the initial stages of a memory recall, but we have a Anot good enough@ detector that keeps us searching our regionally distributed memories until we are confident of our reconstruction of the various scattered parts of the memory. That is, if the situation allows us the time: snap judgements often must be made using still incoherent memories.

Of course, corticocortical coherence in my fiber optic sense is likely to be helpful for preventing incoherent memories. Lapses in neurolinguistic good-enough detectors would produce incoherent sentences, as the neurologists use the word. But we have to be aware of the dangers of using Acoherence@ to make analogies between different levels of organization: I=ve used it at the level of spatiotemporal patterns, Hebb=s cell assembly. You can have incoherence at higher levels B say, for memory recall B for reasons other than corticocortical incoherence. Indeed, at a high level, we have a special term for an incoherent result: we Amix metaphors.@ But such writers are unlikely to have suffered a momentary lapse in their coherence at the corticocortical level.

By the way, Derek, you missed an interesting conversation at dinner. At our end of the table, Sontag and I got to talking with the Chinese scholars who are studying the effect of literary translations on modern culture, and we wandered into the problem of translating multiple levels of meaning, how a word-for-word translation would leave a Chinese reader of Joyce=s opening words in Ulysses missing most of his meaning:

AStately, plump Buck Mulligan came from the stairhead, bearing a bowl of lather on which a mirror and a razor lay crossed. A yellow dressinggown, ungirdled, was sustained gently behind him by the mild morning air. He held the bowl aloft and intoned: Introibo ad altare Dei.@

What=s a poor translator to do? Translating the Latin into local equivalents of AI shall go into the altar of God@ gains a little, yet how many readers in China are likely to recognize that it=s not only the Catholic mass but a parody of it, tinged with blasphemy? So much of our intellectual task, not just in reading Joyce but in interpreting much of everyday conversation, is to locate appropriate levels of meaning between the concreteness of objects and the various levels of category, relationships, and metaphor. Then often, as when reading Joyce, needing to understand things on multiple levels at the same time.

One of the things that I like best about neocortical versions of a Darwin Machine is that I can easily imagine parallel competitions going on, one at the level of the physical setting (piecing together an old Martello gun tower overlooking Dublin Bay with a full of himself medical student about to shave) and another sorting out candidates at the more abstract level of metaphor (ceremonial words and a deliberate pace B but ungirdled gown and an offering of lather?).

DB: So how do we unite the two interpretations? Another level of abstraction, a superior Darwin Machine fed by both outcomes?

WHC: Not necessarily. Your attention might simply alternate between the two levels, just as it does when driving a car and carrying on a conversation at the same time. If there aren=t alternative meldings to consider, that ought to suffice. You don=t need Darwin Machines for everything, just novel tasks with a lot of ambiguity.

Indeed, as a task becomes more familiar, the brain handles it differently; for example, an unfamiliar rapid arm movement might utilize the prefrontal cortex on first acquaintance but, five hours later, might primarily seem to utilize the premotor cortex, cerebellum, and parts of the parietal lobe. I suspect that we often find ways to shortcut, using a prefrontalish Darwin Machine approach only when there isn=t a more familiar routine to invoke.

I=ll bet that analyzing sentences has a number of shortcuts as well, when the subject matter is quite familiar. They might even be more important than the Darwinian competitions, just formed as shortcuts to the Darwinian results from earlier in life.


Notes and References for this chapter

Copyright 2000 by William H. Calvin and Derek Bickerton

The nonvirtual book is available from or direct from MIT Press.

 Email Calvin  

 Email Bickerton  

 Book's Table of Contents  

  Calvin Home Page